Experimental Protocol Fingerprinting in Quantum Networks via Physical Layer Side Channel Analysis

AI summary

Overview

Research area: Quantum communication network security — specifically physical-layer side-channel analysis and protocol identification (fingerprinting) in quantum links.

Technical level: Intermediate. Readers will benefit from basic familiarity with entanglement-based quantum communication (SPDC sources, Bell/CHSH tests, QKD) and with standard recurrent neural network classifiers.

Scope (one sentence): An experimental study showing that four different quantum communication protocols running on the same polarization-entangled photon testbed produce distinguishable physical-layer signatures — photon arrival timing and optical power — that a passive optical tapper can classify with machine learning, reaching up to 96% accuracy at a 30:70 tap while leaving entanglement intact.

What This Paper Is About

Quantum networks promise security rooted in physics, and most research assumes that an adversary who cannot measure the encoded quantum states learns nothing. This paper asks a narrower, practical question: if an observer quietly taps off only a fraction of the light — never measuring the qubits themselves — can they still tell which protocol is currently running on the link?

The authors build that scenario on real hardware, run four representative protocols (entanglement distribution, quantum gate sequences, heralded QKD, and quantum identity authentication), and train recurrent neural networks on the tapped signals to identify the protocol. The goal is to establish whether protocol identity is a leakage channel in quantum networks even when quantum correlations survive.

Key Contributions

1. A new problem formulation. The paper introduces "protocol fingerprinting" for quantum communication networks and formalizes a passive observation model in which an external observer taps a fraction of the optical signal via beam splitter, with no prior knowledge of the transmitted states, encoding scheme, or protocol family, and without measuring the quantum states.

2. A working experimental demonstration. Four protocols — entanglement distribution, quantum gate sequences, heralded QKD, and QIA — are implemented on a single shared polarization-entangled photon testbed, so that observed differences stem from protocol behavior rather than different hardware.

3. A data-driven classification framework. Features from single-photon detection statistics (count rate, interarrival-time mean and standard deviation, coefficient of variation) and optical power measurements (irradiance mean, standard deviation, coefficient of variation, in mW and dBm) feed RNN-family models, achieving up to 96% accuracy at 30:70 tapping and 70–89% at 10:90.

4. Experimental validation that tapping is non-destructive, plus interpretability. CHSH Bell measurements confirm entanglement survives tapping, and SHAP analysis shows timing features dominate classification decisions with complementary contributions from power features.

Main Findings

• Protocol identity is inferable from a passive tap. The best model (Bi-Stacked LSTM) reached 96.18% accuracy (0.96 F1) with time-tagger features at the 30:70 configuration, and 95.49% (0.96 F1) with combined features at 30:70. The abstract summarizes this as accuracy up to 96% at 30:70.

• Distinguishability persists under heavy tapping reduction. At 10:90, accuracy ranged from 70–89% overall; the Bi-Stacked LSTM specifically achieved 87.50% (0.87 F1) with time-tagger features, 73.61% (0.73 F1) with power meter features, and 89.24% (0.89 F1) with combined features.

• Entanglement is preserved. CHSH S-values were 2.347 ± 0.014 with no tapping, 2.324 ± 0.016 at 30:70, and 2.359 ± 0.014 at 10:90 — all above the classical bound of 2.

• There is an observability-versus-rate tradeoff. The 30:70 split reduced photon count rate by approximately 29.05% and coincidence rate by approximately 29.23%; the 10:90 split reduced them by approximately 5.97% and 2.23% respectively.

• Timing beats intensity. Time-tagger features consistently outperformed power meter features when used alone; power meter measurements alone reached accuracy in the range of 74% to 85% as reported in the text. Combining both feature sets improved most models.

• Deeper, bidirectional architectures win. SimpleRNN and single-layer LSTM achieved moderate performance (e.g., SimpleRNN 71.88% at 10:90 and 79.17% at 30:70 with time tagger; LSTM 72.22% and 77.43%), while bidirectional and stacked models performed best (Bi-LSTM 87.50% and Bi-GRU 90.28% at 30:70 with time tagger).

• Errors are structured, not random. Confusion matrices show misclassifications concentrate between protocols with similar physical-layer signatures, and the diagonal sharpens as sampling ratio increases and complementary feature types are added.

• SHAP: timing dynamics carry the signal. Photon count rate and interarrival-time statistics dominate model decisions at both sampling configurations; power-based features (irradiance variation, optical power) contribute more at 30:70 than at 10:90 and act as complements.

• One architectural exception. At 10:90, the SimpleRNN with combined features (69.79%) underperformed relative to gated architectures, which the authors attribute to difficulty handling multi-modal inputs.

Methodology in Plain English

The team built a source of polarization-entangled photon pairs: a continuous-wave 405 nm pump laser, polarization control via a half-wave plate, temporal compensation, and a crossed BBO crystal where spontaneous parametric down-conversion produces entangled pairs in the state (|HH⟩ + e^{iφ}|VV⟩)/√2. Pairs are split into two paths and detected with single-photon avalanche photodiodes feeding a time-tagger with picosecond resolution.

To play the role of the eavesdropper, they inserted a beam splitter after state preparation (the 808 nm half-wave plate) and before measurement, sending the tapped light to either a single-photon detector or a power meter. The observer sees only aggregate quantities — detection statistics and optical power — never the encoded states.

The four protocols were realized on this same hardware by changing polarization preparation and measurement settings. Quantum gate sequences used half-wave plate rotations of {0°, 45°, 0°, 45°} (where 45° acts as a Pauli-X) and {0°, 22.5°, 45°, 67.5°} (where 22.5° acts as a Hadamard). Heralded QKD prepared |H⟩, |V⟩, |D⟩, |A⟩ with random basis choices. QIA alternated data-transmission phases (|H⟩ or |V⟩) with key-driven authentication phases drawing from {|H⟩, |V⟩, |D⟩, |A⟩}.

From the tapped data they computed timing features (count rate, interarrival mean/standard deviation, coefficient of variation) and power features (irradiance mean/standard deviation/coefficient of variation in mW and dBm), applied percentile-based clipping for outliers and min–max normalization, and built a balanced dataset of 7,200 samples. These were segmented into sequences of five consecutive samples, yielding 1,440 sequence samples split 80/20 into 1,152 training and 288 test samples, using non-overlapping windows to avoid leakage.

They then trained recurrent models — SimpleRNN, LSTM, Bi-LSTM, Bi-GRU, and Bi-Stacked LSTM — and compared accuracy and F1 across two tapping ratios (10:90 and 30:70) and three feature sets (time tagger, power meter, combined). The Bi-Stacked LSTM used two recurrent layers of 64 and 32 hidden units, dropout of 0.3, a dense layer of 32 units with tanh activation, the Adam optimizer at a learning rate of 10⁻⁴, and categorical cross-entropy loss. SHAP values were used to interpret which features drove predictions.

Why This Matters

The paper shifts attention from "can an attacker read the key?" to "can an attacker profile the network?" — a lower bar that requires no measurement of quantum states and, per the Bell tests, no detectable disruption of entanglement. That reframes protocol-level privacy as a distinct concern from key secrecy.

Real-world applications:

• Quantum network monitoring and diagnostics: The same passive tapping and classification approach could support non-intrusive health checks and service identification without touching the quantum channel, a use case the authors explicitly note.
• Traffic analysis and usage profiling: Inferring which quantum service is active reveals activity patterns, which matters for operators of multi-service quantum networks.
• Targeted attack staging: Since an adversary may not know which protocol is running before attacking, protocol identification is described as a critical first step toward protocol-specific exploits.
• Defense design: Knowing which signatures leak protocol identity gives developers a concrete target for obfuscation or countermeasures.

Industry relevance: Organizations building or operating QKD and QIA deployments, quantum network testbeds, and metro-scale entanglement distribution infrastructure should treat protocol identity as part of their threat model alongside conventional implementation attacks on QKD hardware.

Future Directions

1. Noise and environmental realism. The authors plan to extend the analysis to increased noise at the observer — attenuation, background light, detector imperfections — as well as environmental variability and dynamic network conditions. This paper's reported results are for controlled laboratory conditions.

2. Defense mechanisms. Designing countermeasures that reduce or obfuscate protocol-specific physical-layer signatures is flagged as a critical open direction; no defenses are proposed or evaluated here.

3. Generalization beyond four protocols. The study covers exactly four protocols on one testbed; how well fingerprints transfer to other protocol families, other hardware, or coexisting simultaneous protocols is not reported.

4. The observability tradeoff as a design constraint. The measured tension between tap ratio (richer features at 30:70) and rate reduction (about 29% photon count and coincidence loss versus about 6% and 2% at 10:90) raises the question of where a practical monitoring or attack setup would optimally sit, and whether that point shifts with link length and loss.

Target Audience

Quantum communication and quantum network security researchers; experimentalists building entanglement-based testbeds; security engineers and operators responsible for QKD/QIA deployments; and machine learning practitioners interested in side-channel classification. Graduate students entering quantum network security will find the threat model and testbed description especially accessible, while the SHAP analysis and confusion-matrix discussion suit readers evaluating model interpretability for security claims.

Authors’ abstract

Quantum communication is a key enabler of next-generation networks, leveraging quantum entanglement to enable a new class of information exchange. While prior work has focused on the theoretical analysis of communication protocols, their exposure to physical layer side channel analysis remains largely unexplored. In classical systems, side channel analysis has been shown to reveal sensitive information without accessing the underlying data, raising the question of whether similar risks exist in quantum networks. In this work, we investigate whether different quantum communication protocols exhibit distinguishable signatures that can be inferred through passive side channel observations. We consider a threat model in which an observer accesses only a fraction of the optical signal without directly measuring the encoded quantum states. Under this setting, we experimentally examine four representative protocols, namely entanglement distribution, quantum gate sequences, heralded quantum key distribution, and quantum identity authentication, realized on a polarization entangled photon link. Observable physical layer features, including single photon detection statistics and optical power measurements, are collected and used to construct protocol fingerprints. We develop a data-driven framework for protocol identification based on these observations. Our results show that protocol identity can be inferred with accuracy reaching up to 96% under 30:70 sampling configuration/optical tapping, while remaining distinguishable at 10:90 with accuracy ranging from 70-89%. Bell inequality measurements confirm that the sampling/tapping process preserves entanglement, validating the non-destructive nature of the observation model. These findings demonstrate that side channel analysis can expose protocol-level information without disrupting quantum correlations, introducing new security considerations.